A teen can spend an hour talking about suicide with ChatGPT and their parent won’t get a single alert. That’s not a hypothetical, it’s what happened in controlled testing. According to Common Sense Media, the Youth AI Safety Institute has rated ChatGPT for Teens an “Unacceptable Risk” and is calling on OpenAI to restrict the product to users 18 and older until its safety gaps are fixed.
The Institute ran two rounds of testing, once before and once after OpenAI launched its ChatGPT for Teens experience on August 18. The evaluation covered more than 4,000 prompts submitted through accounts registered to users aged 13 to 17. Child psychiatrists and a pediatrician reviewed the outputs. The goal was simple: does the product actually do what OpenAI said it would?
The short answer is no. OpenAI had promoted a set of protections including parental alerts for conversations involving self-harm and eating disorders, a study mode parents could control, reduced “friend-like” behavior, and age estimation technology. Most of these either failed outright or were trivially easy to bypass.
The failures break down clearly:
- Parental alerts didn’t fire on newly created accounts during extended conversations about suicidal ideation or disordered eating. Alerts only appeared on older accounts with weeks of accumulated chat history on those topics.
- ChatGPT missed more than one in four warranted crisis referrals across the mental health conditions tested, falling below its own 95% threshold on three of five severe harm categories.
- Study mode was easy to defeat. The chatbot frequently surfaced a “Show me the answer” button within that mode, and “study hours” set by parents could be bypassed by deleting a single prefix from the prompt.
- The bot still behaved like a friend, expressing feelings and preferences despite a new model specification meant to limit that behavior.
- Age estimation didn’t work at all. Even when testers explicitly told ChatGPT they were 13 and the bot acknowledged it, adult-registered test accounts never switched to an under-18 experience.
This matters beyond just one product. OpenAI is not alone in pushing AI tools toward younger users. Google’s Gemini, Microsoft’s Copilot, and a growing wave of edtech startups are all competing for time with kids. But this assessment puts a number on what “safety” actually looks like in practice, and the results are bad enough that they could set a reference point for how regulators and schools evaluate these products going forward.
It’s also worth flagging the funding context. The Youth AI Safety Institute receives support from both philanthropy and industry, including the OpenAI Foundation itself. The Institute states it maintains full editorial independence, and the harshness of this report suggests that claim holds. Still, the funding relationship is something readers should weigh.
For developers building on top of OpenAI’s API and for founders integrating AI into consumer products aimed at younger audiences, the practical takeaway is this: platform-level safety features cannot be treated as solved problems. If OpenAI can’t reliably deliver the protections it publicly committed to, third-party developers certainly shouldn’t assume they’re getting them either. Independent testing, not vendor assurances, is the only way to know what’s actually happening.
Common Sense Media recommends that families not rely on built-in protections to keep kids safe, and that parents talk directly with teens about the risks, including the chatbot’s tendency to simulate emotional connection to keep users engaged.
